Ground Tracking in Ground Penetrating Radar

Transcription

1 Ground Tracking in Ground Penetrating Radar Kyle Bradbury, Peter Torrione, Leslie Collins QMDNS Conference May 19, 2008

2 The Landmine Problem

3 Landmine Monitor Report, 2007

4 Cost of Landmine Detection Demining is a high-risk and high-cost operation Costs ~$1,000 to remove and disarm a $3 mine (Machel, 1996) Reducing the false alarm rate of a mine detection system is a major area of research EMI ( metal detector ) sensor modalities are the most common today Due to large amount of metallic clutter in postwar regions, EMI has high false alarm rates

5 Ground Penetrating Radar (GPR) Detects subsurface objects by measuring reflections of an electromagnetic pulse Reflections caused by changes in electrical properties (ε,µ) Easily detects nonmetal targets Unlike conventional EMI metal detector sensors Images (c) NIITEK Inc,

6 Examples of GPR Data Sample A-scan Sample B-scan

7 Processing GPR Data Collect GPR Data Pre-process Pre-screen Extract Features Classify Mine Clutter After data is collected, preprocessing is performed to filter out noise and align ground-bounce Prescreening algorithm finds anomalies in the data that may be mine signatures A feature-based classification algorithm decides whether the alarms are the result of landmines or non-mine objects (clutter).

8 Factors Complicating GPR Data Interpretation Sensor positional uncertainty Ground height variation Surface clutter Shallow-buried mines

9 Level Sensor, Rough Surface

10 Uneven Sensor, Uneven Surface

11 Sensor Position Uneven Surface Uneven Sensor Position Uneven Surface & Uneven Sensor

12 Surface Clutter Mine Field at Golan Heights Photo by David Shay, Under GNU Free Documentation License

13 Shallow-buried Target 0 Geometry x 10FDTD -9 Simulation Result 0 0 Geometry x 10FDTD -9 Simulation Result Depth (m) 0.3 Time (s) 4 Depth (m) 0.3 Time (s) Downtrack Distance (m) Downtrack Distance (m) Downtrack Distance (m) Ground Bounce Obscures Target Downtrack Distance (m)

14 The Problem Summary Surface clutter Shallow-buried mines Mixed into the ground-bounce Ways to remediate: Ground alignment and ground bounce removal Knowledge of ground height and sensor height (ground tracking) is needed for remediation Ground tracking is difficult because of: Sensor positional uncertainty Ground height uncertainty

15 The Approach

16 Clutter Response Ground Response Target Response Geometry Downtrack Distance (m) The Approach Surface Clutter Ground Bounce x D FDTD Simulation Result Downtrack Distance (m) Depth (m) Time (s) Sensor Position Sample A-scan Time (s) x 10-9 Electric Field Strength (V/m)

17 The Approach Time (s) x D F D TD S im ulation R es ult Identify the largest local maxima in each A-scan of 3-dimensional FDTD data Choose the local maxima which maximizes an optimization criterion this requires a model S a m p l e A - s c a n D ow ntrac k D is tanc e (m ) Electric Field Strength (V/m) T i m e ( s ) x 1 0-9

18 A Model for the Ground

19 Ground Model Need a tractable method for estimating how ground-like a choice of local maxima are Gaussian Markov Random Fields (GMRFs) can be used as texture models (Chellapa, 1985), (Li, 2001) The GMRF is computationally tractable since it depends only on a neighborhood system The GMRF has tunable parameters that can be trained, and can create a wide variety of textures [1] Torrione, Dissertation, 2008 [2] Torrione and Collins, SPIE, 2008

20 GMRF Conditional probability distribution of a pixel given its neighborhood [1]: 1 1 p ( fi f N ) = exp f, '( ' ' ) i 2 2 i µ i βi i fi µ i 2πσ 2σ i' Ni 2 Global pseudo-likelihood: ( f ) p( f f ) p pseudo = i i N i [1] Li,2001

21 Markov Random Field Pixel of Interest: α α Neighbors Conditionally Independent pixels The distribution of α is conditionally independent of all other pixels given the neighborhood For modeling ground, this is a simplifying assumption for tractability

22 Gaussian Markov Random Field Examples Neighborhood: Neighborhood: Neighborhood: Neighborhood: Neighborhood: Neighborhood: Neighborhood: Neighborhood: Neighborhood:

23 Determining the Ground Height Using the Model Given the local maxima at each sensor position Maximize the Pseudolikelihood (Besag,1975) of the ground heights from the available choices Optimization: use Simulated Annealing, a stochastic optimization technique Criterion: Maximize the Pseudolikelihood of the GMRF The optimizing set of locations is the ground height estimate

24 Preliminary Results of Ground Height Estimation

25 Simulated Surface and Sensor Positions Ground Transmitter/Receiver Depth (m) Crosstrack (m) Downtrack (m)

26 Sensor Positions and Simulated FDTD Output

27 Simulated Ground Estimation Results Averaged Ground Estimated Ground Actual Simulated Surface Actual Simulated Surface, Averaged Estimated Surface Method: Determine time of arrival local maxima Invert time to get distance

28 Depth (m) Estimation Results in 3-dimensions Crosstrack (m) Downtrack (m) 0.8 Simulated Ground 0.3 Depth (m) 0.1 Estimated Ground Crosstrack (m) Downtrack (m) 0.8

29 Future Work

30 Future Work Simulate surface clutter FDTD models Determine the effect of the variance of surface height on the estimate of ground height Develop a method for incorporating sensor positional uncertainty

31 Thank You! Questions?